Generative Modeling and Computational Photography Researcher | ML/AI Engineer
I'm a 2nd year CS PhD student at the University of Wisconsin-Madison, advised by Professor Mohit Gupta.
I have been fortunate to work with many brilliant professors and experts throughout my journey. Prof. Renu Rameshan ignited my interest in computer vision early during my Bachelor's at IIT Mandi which also led me to exploring generative modeling for outdoor lighting during my junior year with Prof. Jean-François Lalonde and Dr. Yannick Hold-Geoffroy. After graduating, I spent a wonderful year at IIT Madras getting closer to sensors alongside developing 3D reconstruction and generative facial video editing methods as a Bosch Research Fellow at the Computational Imaging Lab directed by Prof. Kaushik Mitra. Interestingly, my vision sensors and generative interests got a human touch to them as I spent the summer of '24 at the Computational Behaviour Lab, advised by Prof. Antti Oulasvirta and now Professor, Yue Jiang.
My research focuses on generative modeling for imaging, vision, and photography. Currently, I'm working on HDR text-to-image models, semantically controllable dynamic range expansion, and lifting diffusion priors to long-horizon text-to-video generation.
Apart from research, I enjoy riding my motorcycle, baking and trekking. I also love to travel during the holidays to meet my close friends Julian Fabinc and Noemi Ippolito.
Adapting large-scale generative priors for photon-counting or SPAD-based neural ISPs to image under extreme motion.
High Dynamic Range 3D Gaussian Splatting from raw images. Achieves night-time HDR 3D reconstructions as well.
Identity-preserving facial video editing with diffusion models.
CVPR 2026 (Highlight; Top 10%)
ToG 2025 & SIGGRAPH 2026
Preprint, 2025
Diffusion for automatic GUI designing & ideation
WACV-W, 2025
Identity-Preserving Facial Video Editing with Diffusion
BMVC, 2024
ICCP, 2024
ICLR Tiny Paper, 2024 (Oral Presentation)
Unconditional face generation at the speed and size of DCGAN using a relatively tiny training dataset (~4000).
ICLR Tiny Paper, 2024 (Oral Presentation)
Synthetic dataset generation and sinusoidal activations lead to a winning solution over the final displacement error.